3 ms·

We can get more transistors by going 3D, but what about the heat? The real tragedy has been the degradation (although not yet death) in Dennard Scaling. And ev

We can get more transistors by going 3D, but what about the heat? The real tragedy has been the degradation (although not yet death) in Dennard Scaling. And even then Landauer's limit will begin to rear it's head around 10,000-100,000X lower power. Reversible computing is an option, but does not cover I/O and will furthermore set us back in performance and transistor efficiency.

posix_compliant 7y ago
If memory serves correctly, about a year ago there was a shift in Bitmain Bitcoin miners towards a 3D chip architecture that was fairly effective; it might be a domain specific application, though.
NotSammyHagar 7y ago
Had to look up those two terms, from Wikipedia: Dennard scaling, also known as MOSFET scaling, is a scaling law based on a 1974 paper co-authored by Robert H. Dennard, after whom it is named.[1] Originally formulated for MOSFETs, it states, roughly, that as transistors get smaller, their power density stays constant, so that the power use stays in proportion with area; both voltage and current scale (downward) with length. Landauer's principle is a physical principle pertaining to the lower theoretical limit of energy consumption of computation. ... At 20 °C ..., the Landauer limit represents an energy of approximately 0.0175 eV, or 2.805 zJ. Theoretically, room‑temperature computer memory operating at the Landauer limit could be changed at a rate of one billion bits per second with energy being converted to heat in the memory media at the rate of only 2.805 trillionths of a watt (that is, at a rate of only 2.805 pJ/s). Modern computers use millions of times as much energy per second. [But there's a limit to how far we can improve that leading to hitting it around 2050].
- deepnotderp 7y ago
  Modern computers using millions more is a little misleading, since a lot of that goes to architectural overhead and other things than single bit computations.

posix_compliant 7y ago

If memory serves correctly, about a year ago there was a shift in Bitmain Bitcoin miners towards a 3D chip architecture that was fairly effective; it might be a domain specific application, though.

NotSammyHagar 7y ago

Had to look up those two terms, from Wikipedia: Dennard scaling, also known as MOSFET scaling, is a scaling law based on a 1974 paper co-authored by Robert H. Dennard, after whom it is named.[1] Originally formulated for MOSFETs, it states, roughly, that as transistors get smaller, their power density stays constant, so that the power use stays in proportion with area; both voltage and current scale (downward) with length. Landauer's principle is a physical principle pertaining to the lower theoretical limit of energy consumption of computation. ... At 20 °C ..., the Landauer limit represents an energy of approximately 0.0175 eV, or 2.805 zJ. Theoretically, room‑temperature computer memory operating at the Landauer limit could be changed at a rate of one billion bits per second with energy being converted to heat in the memory media at the rate of only 2.805 trillionths of a watt (that is, at a rate of only 2.805 pJ/s). Modern computers use millions of times as much energy per second. [But there's a limit to how far we can improve that leading to hitting it around 2050].

deepnotderp 7y ago
Modern computers using millions more is a little misleading, since a lot of that goes to architectural overhead and other things than single bit computations.